feat: dynamic tool registration — agent creates new tools at runtime

[nv78]

What this does

The agent can now build its own tools on the fly. When it encounters a task no existing tool covers, it calls register_tool — writes the Python implementation, defines the schema, and the tool becomes available for the rest of the conversation.

The loop:

1. Agent hits an unsupported task → calls think to plan the implementation
2. Calls register_tool(name, description, parameters_schema, python_code)
3. Backend validates + compiles the code in a sandboxed namespace
4. Tool appears in the LLM's tool list on the next iteration
5. Agent calls its new tool as if it were built-in

Architecture

Session Registry

Each conversation gets a session_registry = {specs: [], executors: {}}. This mutable dict is threaded through every layer — _run → _run_anthropic/_run_openai → _execute_tool — so dynamic tools are injected into every subsequent LLM call within the same conversation.

Safety (AST-based validation)

_validate_tool_code() walks the AST before any exec():

• Blocks imports of os, sys, subprocess, socket, shutil, ctypes, pickle, etc.
• Blocks dangerous builtins: eval, exec, open, compile, __import__
• Blocks dunder attribute access (__globals__, __subclasses__, etc.)
• Requires a run(inputs: dict) -> str function to be defined

Allowed libraries in the sandbox: json, re, math, datetime, collections, itertools, urllib, requests, pandas, numpy, bs4.

Execution order

register_tool always runs sequentially (it mutates session_registry). All other tools continue to run in parallel via ThreadPoolExecutor.

Frontend

• tool_registered SSE event → new reasoning step with emerald styling + wrench icon
• Shows the fn name(inputs) signature and description so the user can see what was built

Example use cases

• "Fetch live stock prices for AAPL, MSFT, GOOGL" → agent creates fetch_stock_price wrapping the Yahoo Finance API, then calls it 3 times in parallel
• "Parse all invoice totals from this text" → agent creates parse_invoice with regex logic, then calls it
• "Convert this CSV to a markdown table" → agent creates csv_to_markdown using pandas, then calls it
• "Query the OpenWeatherMap API for Berlin" → agent builds the wrapper once, reuses it for multiple cities

Test plan

• Ask "what's the weather in Tokyo?" (no weather tool exists) → agent should call register_tool to build a weather API wrapper, then call it
• Ask to parse a custom data format → agent should register a parser and call it
• Try to register a tool using import os → should be blocked with a validation error
• Try to register a tool without a run function → should return error
• Register a tool, then ask a follow-up that uses it → tool should persist across turns
• Verify tool_registered event appears in the reasoning panel with emerald color

https://claude.ai/code/session_01C9mHttiQ4ZAaBbQecVV7uu